TimesFM Forecasting

What it does

This skill wraps Google Research's TimesFM 2.5 (200M-parameter time-series foundation model) for agent-friendly local inference. Feed in any univariate series and get back median point forecasts plus 10 quantile bands (q10–q90) for calibrated prediction intervals — all zero-shot.

• Accepts CSV, pandas DataFrame, or NumPy array inputs
• Context windows up to 16,384 points; efficient batch forecasting of many series
• Covariate support (price, promotions, holiday flags, region) via forecast_with_covariates()
• Lightweight anomaly detection using quantile intervals
• Mandatory check_system.py preflight that verifies RAM, VRAM, disk, and Python version before any model download

Who it's for

• Analysts tired of hand-tuning ARIMA/ETS parameters
• Ops and planning teams needing fast demand or revenue forecasts
• Engineers handling sensor, energy, or vitals telemetry at scale
• Researchers who must report forecast uncertainty, not just point estimates

Not the right tool for coefficient-level statistical inference, series classification/clustering, or multivariate causality tests.

Example uses

1. 12-month sales forecast: load monthly_sales.csv and emit a results CSV with forecasts plus 80% intervals for every product column.
2. Sensor anomaly watch: forecast the next 24 hours from the last 1,000 points and flag actuals outside the 90% interval as "Critical".
3. Baseline comparison: compute MAE for TimesFM vs. statsmodels ARIMA on the same holdout window to justify model choice.

Install the skill I found on Claude Skill Mart.
Copy the skills/timesfm-forecasting folder from the GitHub repo K-Dense-AI/scientific-agent-skills into my ~/.claude/skills/timesfm-forecasting/.
When it's done, tell me in one line what this skill can do.